paper-summarizer

name	paper-summarizer
description	Summarize one local paper workspace without network access. Use existing parse artifacts and metadata to write, assemble, and validate `draft.md`.

Use this skill to turn parsed paper content into a publishable summary draft.

Path rule:

• references/... and scripts/... are relative to this skill directory.
• From repo root, run python skills/paper-summarizer/scripts/.... Repo-level configs stay under config/.

Load only these files when needed:

• config/summary_template.json
• config/embodied_ai_terminology.json for embodied AI / robot learning terminology normalization
• references/artifact-contract.md
• references/summarize_status.schema.json

Use these scripts for deterministic steps:

• skills/paper-summarizer/scripts/write_summarize_status.py
• skills/paper-summarizer/scripts/validate_summary.py

Inputs:

• data/papers/<paper_id>/parse/main_text.md
• data/papers/<paper_id>/parse/figure_index.md
• data/papers/<paper_id>/parse/figure_index.json
• data/papers/<paper_id>/parse/figures/
• data/papers/<paper_id>/parse/status.json
• data/papers/<paper_id>/acquire/metadata.json

Work:

1. summarize_text
2. summarize_figures
3. assemble_summary
4. validate

Treat the workflow as complete only when summarize/status.json.status == "succeeded".

Use these commands and outputs:

1. Summarize text:
   • Read main_text.md and config/summary_template.json
   • Write summary_text.md
   • Follow config/summary_template.json for article-type-specific structure and writing guidance.
   • Include only formulas or code/pseudocode that are necessary to explain the method's core mechanism, inline with the relevant body text.
   • For embodied AI / robot learning papers, use config/embodied_ai_terminology.json only to normalize terms that appear in the paper.
2. Summarize figures:
   • Read figure_index.md, figure_index.json, and selected images only
   • Select at most 2 to 3 figures
   • Write selected_figures.json and figure_notes.md
3. Assemble summary:
   • Assemble from summary_text.md, figure_notes.md, and local metadata
   • Follow config/summary_template.json
   • Put available metadata fields directly under the top title as part of the summary template
   • Include header_fields. Omit optional_header_fields when the value is missing, empty, not_specified, or unknown
   • Keep figures inline near relevant text, using relative image paths from summarize/draft.md such as ../parse/figures/figure-01.png
   • Write draft.md and status.json
4. Validate:
   • Run python skills/paper-summarizer/scripts/validate_summary.py <summarize-dir>/draft.md
   • Then run python skills/paper-summarizer/scripts/write_summarize_status.py data/papers/<paper_id>

Complete only when:

• summarize/status.json.status == "succeeded"

Fail when:

• a required input file is missing
• a step status file reports failure or invalid output
• summarize/status.json.status is failed or blocked

Guardrails:

• This skill is local-only.
• Deterministic work uses scripts. Judgment work stays in the current Codex session.
• Each step should read only its declared inputs and status files.
• Do not launch nested codex exec.
• Stop immediately when the current step fails.